Eisenberg Einklang

[EisenbergAudio]

OK - I thought that the existing sounds were resynthesized versions of samples, and thus perhaps it was possible to resynthesize your own samples to include as morphable tone models.

Hi bharris22,

The engine uses a whole bunch of samples for building each model. So you always need a group of samples from the same source. For each tone color from the normal tone packs (Red, Green, Blue) we've used around 10 to 20 samples for the modeling. For each tone color from the premium packs we've used 30 to 60 samples as they use heavy intra-morphing. That is one tone maybe sounds like instrument A when played piano in the bass section and smoothly morphs to instrument B in the high range but sounds like instrument C over the whole range when played forte.

Cheers
Gunnar

[bharris22]

Thank you, Gunnar. Is it possible to include user-created tone models for a future update (providing the user could provide a group of samples from the same source)?

[EisenbergAudio]

Thank you, Gunnar. Is it possible to include user-created tone models for a future update (providing the user could provide a group of samples from the same source)?

That is definitely our goal but at the moment the whole training is part of a "not so handy" process. Meaning you have to have some expert knowledge on the AI training. We hope to find a solution for simplifying this in the future.

[Krakatau]

Oops. i talked too fast, i'm afraid. saying "model" didn't suggest the use of samples

Sorry...

[EisenbergAudio]

Oops. i talked too fast, i'm afraid. saying "model" didn't suggest the use of samples

Sorry...

Well, what you've said isn't wrong. We somewhat enter a philosophic area here You will not find any sample material in the tone color models, neither in pcm format nor in some direct transformation. It is that the AI learns the structures of the samples and recognizes the underlying patterns during the training, i.e. during sound design. So the information is still in there but it is scattered all over the structure of the AI and it's not human readable anymore. Further it's not trivial telling before training which information of the samples is kept and which is thrown away for good. So the AI also performs a data compression and reduction.

In the end, what we think of for the future might be a hybrid engine, allowing users to load their samples and let them get analyzed like in traditional additive synthesizers. The engine would guess from which underlying source this sound comes and apply the specific model parameters that match the source. E.g. the user loads a specific violin sample, the engine analyses it, recognizes the violin model with the closest matching sound properties and applies it. Then the engine knows how this specific sample will sound when played louder, softer or changes pitch and modulation.

The user could of course also change the applied model wich would result in interesting cross synthesis instruments... sky is the limit. But that all means a lot more research effort and it will take time to implement it.

Cheers
Gunnar

[Krakatau]

Oops. i talked too fast, i'm afraid. saying "model" didn't suggest the use of samples

Sorry...

Well, what you've said isn't wrong. We somewhat enter a philosophic area here You will not find any sample material in the tone color models, neither in pcm format nor in some direct transformation. It is that the AI learns the structures of the samples and recognizes the underlying patterns during the training, i.e. during sound design. So the information is still in there but it is scattered all over the structure of the AI and it's not human readable anymore. Further it's not trivial telling before training which information of the samples is kept and which is thrown away for good. So the AI also performs a data compression and reduction.

In the end, what we think of for the future might be a hybrid engine, allowing users to load their samples and let them get analyzed like in traditional additive synthesizers. The engine would guess from which underlying source this sound comes and apply the specific model parameters that match the source. E.g. the user loads a specific violin sample, the engine analyses it, recognizes the violin model with the closest matching sound properties and applies it. Then the engine knows how this specific sample will sound when played louder, softer or changes pitch and modulation.

The user could of course also change the applied model wich would result in interesting cross synthesis instruments... sky is the limit. But that all means a lot more research effort and it will take time to implement it. aèèée silicon/Big Sur compatibility

Cheers
Gunnar

Le me bump this thread for a request

Gunnar, do you ever plan to update einklang for Apple silicon/big sur compatibility ?
too bad because the M1 very likely has enough power to run einklang flawlessly at his most demanding performances

I'll be glad if it has a chance to happen, you made a very unique instrument in its area !!

[EisenbergAudio]

Thanks Krakatau. At the moment we have no plans for upgrading it but I would never say never. When we started Einklang we had tons of ideas for improvements and even for a bigger redesign that would much improve the natural sounds... Yes, the performance that is available nowadays would help a lot, but at the moment the core team that made Einklang is busy with other things. Let's see, what the future brings.

[aMUSEd]

Oops. i talked too fast, i'm afraid. saying "model" didn't suggest the use of samples

Sorry...

Well, what you've said isn't wrong. We somewhat enter a philosophic area here You will not find any sample material in the tone color models, neither in pcm format nor in some direct transformation. It is that the AI learns the structures of the samples and recognizes the underlying patterns during the training, i.e. during sound design. So the information is still in there but it is scattered all over the structure of the AI and it's not human readable anymore. Further it's not trivial telling before training which information of the samples is kept and which is thrown away for good. So the AI also performs a data compression and reduction.

In the end, what we think of for the future might be a hybrid engine, allowing users to load their samples and let them get analyzed like in traditional additive synthesizers. The engine would guess from which underlying source this sound comes and apply the specific model parameters that match the source. E.g. the user loads a specific violin sample, the engine analyses it, recognizes the violin model with the closest matching sound properties and applies it. Then the engine knows how this specific sample will sound when played louder, softer or changes pitch and modulation.

The user could of course also change the applied model wich would result in interesting cross synthesis instruments... sky is the limit. But that all means a lot more research effort and it will take time to implement it.

Cheers
Gunnar

Sounds a lot like the Hartmann Neuron (the model maker in particular)